"""Reduce training padding while retaining sequential validation batches.""" from torch.utils.data import SequentialSampler from trl import SFTTrainer class OptimizedSFTTrainer(SFTTrainer): def _get_eval_sampler(self, eval_dataset): # Training length grouping should not change validation's batch ordering. # Accelerate shards this sequential loader and gathers the full split. return SequentialSampler(eval_dataset)